home-ops: Skill for Claude Code

Intel GPUs: device plugins, VA-API and telemetry

Relocated verbatim from AGENTS.md on 2026-09-01 so it loads only when this subsystem is in play. The text below is unchanged; only line breaks were inserted. AGENTS.md keeps a one-sentence pointer. Add new findings here or to the owning document, not back into AGENTS.md - see its "Maintaining this file" section for the rule.

• NEVER rename a DRM device node with generic-device-plugin's mountPath, and never ship a GPU change without running the VA-API check. The b70 group remaps the Arc B70 to card0/renderD128; that rename is fatal to VA-API, because libdrm ignores the path you pass, fstat()s the fd, reads /sys/dev/char/<major>:<minor>/uevent and reopens the canonical DEVNAME it finds there (dri/renderD129) - a path the container does not have. vaGetDisplayDRM() then fails before any driver loads. Level Zero (vllm, comfyui) opens whatever /dev/dri/render* exists and is unaffected, so the AI stack stays green while transcoding is completely dead; that asymmetry hid a total 3-day Tdarr outage (2026-08-26, PR #1443). VA-API consumers must use devic.es/b70-vaapi, which exposes the same card under its kernel names. Two related traps in the same config: its device IDs are sha1(count + every host path in the group), so adding a path to an existing group changes all its IDs and invalidates kubelet's live allocations for pods holding that resource (add a new group instead); and the config is a subPath mount, which kubelet never refreshes, so the configMapGenerator hash must stay enabled or a config-only edit is inert until someone restarts the DaemonSet by hand. Allocatable capacity is not proof that transcoding works - verification commands: docs/media-stack.md "Verifying VA-API after a GPU change"; mechanism and evidence: docs/ai-gpu-changelog.md (2026-08-29).

• No xpu-smi/level-zero/DCGM-equivalent GPU exporter is deployed, so per-engine busy %, VRAM utilization, and clocks are not queryable in Prometheus for either Intel GPU. Verified live 2026-08-26 while building ai/gpu-node-dashboard. What does exist: the discrete Arc B70 (talos-3 only, devic.es/b70) surfaces a kernel xe-driver hwmon chip (node_hwmon_chip_names{chip_name="xe"}, chip id 0000:02:01_0_0000:03:00_0) with temp2=package/temp3=VRAM (node_hwmon_temp_celsius), power1 cap/crit (node_hwmon_power_cap_watt/_crit_watt, static config values, not live draw), fan1 RPM (node_hwmon_fan_rpm), and energy counters energy1=card/energy2=package (node_hwmon_energy_joule_total, rate() them for live watts - there is no direct power-draw gauge). The fleet iGPU (gpu.intel.com/xe, on-die, all 3 nodes) has no dedicated hwmon chip at all - only kube-state-metrics' device-plugin kube_node_status_allocatable/kube_pod_container_resource_requests{resource="gpu_intel_com_xe"} show allocation/usage, no thermal or utilization signal. Dashboard and verified panel list: kubernetes/apps/base/ai/gpu-node-dashboard/app/gpu-node.json.